This collection demonstrates how to construct and train a deep, bidirectional stacked LSTM using CNN features as input with CTC loss to perform robust word recognition.
The model is a straightforward adaptation of Shi et al.'s CRNN architecture (arXiv:1507.0571). The provided code downloads and trains using Jaderberg et al.'s synthetic data (IJCV 2016), MJSynth.
Notably, the model achieves a lower test word error rate (1.82%) than CRNN when trained and tested on case-insensitive, closed vocabulary MJSynth data.
Written for Python 2.7. Requires TensorFlow >=1.10 (deprecation warnings exist for TF>1.10, but the code still works).
The model and subsequent experiments are more fully described in Weinman et al. (ICDAR 2019)
The model as built is a hybrid of Shi et al.'s CRNN architecture (arXiv:1507.0571) and the VGG deep convnet, which reduces the number of parameters by stacking pairs of small 3x3 kernels. In addition, the pooling is also limited in the horizontal direction to preserve resolution for character recognition. There must be at least one horizontal element per character.
Assuming one starts with a 32x32 image, the dimensions at each level of filtering are as follows:
| Layer | Op | KrnSz | Stride(v,h) | OutDim | H | W | PadOpt |
|---|---|---|---|---|---|---|---|
| 1 | Conv | 3 | 1 | 64 | 30 | 30 | valid |
| 2 | Conv | 3 | 1 | 64 | 30 | 30 | same |
| Pool | 2 | 2 | 64 | 15 | 15 | ||
| 3 | Conv | 3 | 1 | 128 | 15 | 15 | same |
| 4 | Conv | 3 | 1 | 128 | 15 | 15 | same |
| Pool | 2 | 2,1 | 128 | 7 | 14 | ||
| 5 | Conv | 3 | 1 | 256 | 7 | 14 | same |
| 6 | Conv | 3 | 1 | 256 | 7 | 14 | same |
| Pool | 2 | 2,1 | 256 | 3 | 13 | ||
| 7 | Conv | 3 | 1 | 512 | 3 | 13 | same |
| 8 | Conv | 3 | 1 | 512 | 3 | 13 | same |
| Pool | 3 | 3,1 | 512 | 1 | 13 | ||
| 9 | LSTM | 512 | |||||
| 10 | LSTM | 512 |
To accelerate training, a batch normalization layer is included before each pooling layer and ReLU non-linearities are used throughout. Other model details should be easily identifiable in the code.
The default training mechanism uses the ADAM optimizer with learning rate decay.
The original CRNN uses a single 3x3 convolution in the first two conv/pool stages, while this network uses a paired sequence of 3x3 kernels. This change increases the theoretical receptive field of early stages of the network.
As a tradeoff, we omit the computationally expensive 2x2x512 final convolutional layer of CRNN. In its place, this network vertically max pools over the remaining three rows of features to collapse to a single 512-dimensional feature vector at each horizontal location.
The combination of these changes preserves the theoretical receptive field size of the final CNN layer, but reduces the number of convolution parameters to be learned by 15%.
Another important difference is the lack of zero-padding in the first convolutional layer, which can cause spurious strong filter responses around the border. By trimming the first convolution to valid regions, this model erodes the outermost pixel of values from the response filter maps (reducing height from 32 to 30 and reducing the width by two pixels).
This approach seems preferable to requiring the network to learn to ignore strong Conv1 responses near the image edge (presumably by weakening the power of filters in subsequent convolutional layers).
We include batch normalization after each pair of convolutions (i.e., after layers 2, 4, 6, and 8 as numbered above). The CRNN does not include batch normalization after its first two convolutional stages. Our model therefore requires greater computation with an eye toward decreasing the number of training iterations required to reach converegence.
The first two pooling stages of CRNN downsample the feature maps with a stride of two in both spatial dimensions. This model instead preserves sequence length by downsampling horizontally only after the first pooling stage.
Because the output feature map must have at least one timeslice per character predicted, overzealous downsampling can make it impossible to represent/predict sequences of very compact or narrow characters. Reducing the horizontal downsampling allows this model to recognize words in narrow fonts.
This increase in horizontal resolution does mean the LSTMs must capture more information. Hence this model uses 512 hidden units, rather than the 256 used by the CRNN. We found this larger number to be necessary for good performance.
To completely train the model, you will need to download the mjsynth
dataset and pack it into sharded TensorFlow records. Then you can start
the training process, a tensorboard monitor, and an ongoing evaluation
thread. The individual commands are packaged in the accompanying Makefile.
make mjsynth-download
make mjsynth-tfrecord
make train &
make monitor &
make test
To monitor training, point your web browser to the url (e.g., (http://127.0.1.1:8008)) given by the Tensorboard output.
Note that it may take 4-12 hours to download the complete mjsynth data
set. A very small set (0.1%) of packaged example data is included; to
run the small demo, skip the first two lines involving mjsynth.
With a GeForce GTX 1080, the demo takes about 20 minutes for the validation character error to reach 45% (using the default parameters); at one hour (roughly 7000 iterations), the validation error is just over 20%.
With the full training data, by one million iterations the model typically converges to around 5% training character error and 27.5% word error.
The evaluate script (src/evaluate.py) streams statistics for one
batch of validation (or evaluation) data. It prints the iteration,
evaluation batch loss, label error (percentage of characters predicted
incorrectly), and the sequence error (percentage of words—entire
sequences—predicted incorrectly).
The test script (src/test.py) tallies statistics, finally
normalizing for all data. It prints the loss, label error, total
number of labels, sequence error, total number of sequences, and the
label error rate and sequence error rate.
To see the output of a small set of instances, the validation script
(src/validation.py) allows you to load a model and read an image one
at a time via the process's standard input and print the decoded
output for each. For example
cd src ; python validate.py < ~/paths_to_images.txt
Alternatively, you can run the program interactively by typing image paths in the terminal (one per line, type Control-D when you want the model to run the input entered so far).
There are many command-line options to configure training
parameters. Run train.py or test.py with the --help flag to see
them or inspect the scripts. Model parameters are not command-line
configurable and need to be edited in the code (see src/model.py).
Dynamic data can be used for training or testing by setting the
--nostatic_data flag.
You can use the --ipc_synth boolean flag [default=True] to determine
whether to use single-threaded or a buffered, multiprocess synthesis.
The --synth_config_file flag must be given with --nostatic_data.
The MapTextSynthesizer library supports training with dynamically synthesized data. The relevant code can be found within MapTextSynthesizer/tensorflow/generator
By default, recognition occurs in "open vocabulary" mode. That is, the system observes no constraints on producing the resulting output strings. However, it also has a "closed vocabulary" mode that can efficiently limit output to a given word list as well as a "mixed vocabulary" mode that can produce either a vocabulary word from a given word list (lexicon) or a non-vocabulary word, depending on the value of a prior bias for lexicon words.
Using the closed or mixed vocabulary modes requires additional software. This repository is connected with a fork of Harald Scheidl's CTCWordBeamSearch, obtainable as follows:
git clone https://github.com/weinman/CTCWordBeamSearch
cd CTCWordBeamSearch
git checkout var_seq_lenThen follow the build instructions, which may be as simple as running
cd cpp/proj
./buildTF.shTo use, make sure CTCWordBeamSearch/cpp/proj (the directory
containing TFWordBeamSearch.so) is in the LD_LIBRARY_PATH when
running test.py or validate.py (in this repository).
This version uses the TensorFlow Dataset for fast I/O. Training, testing, validation, and prediction use a custom Estimator.
Please cite the following paper if you use this code in your own research work:
@inproceedings{ weinman19deep,
author = {Jerod Weinman and Ziwen Chen and Ben Gafford and Nathan Gifford and Abyaya Lamsal and Liam Niehus-Staab},
title = {Deep Neural Networks for Text Detection and Recognition in Historical Maps},
booktitle = {Proc. IAPR International Conference on Document Analysis and Recognition},
month = {Sep.},
year = {2019},
location = {Sydney, Australia}
}
This work was supported in part by the National Science Foundation under grant Grant Number 1526350.